WebGPU

1. 引言

本节为非规范性内容。

图形处理单元（GPU）在个人计算领域中已成为实现丰富渲染和计算应用的关键。WebGPU 是一个为 Web 暴露 GPU 硬件能力的 API。该 API 从零开始设计，能够高效映射到（2014年之后的）原生 GPU API。WebGPU 与 WebGL 无关，也不直接针对 OpenGL ES。

WebGPU 将物理 GPU 硬件视为 GPUAdapter。它通过 GPUDevice 提供与适配器的连接，设备负责管理资源，并通过设备的 GPUQueue 执行命令。GPUDevice 可能拥有自己的高速存储器以供处理单元访问。GPUBuffer 和 GPUTexture 是由 GPU 内存支持的物理资源。GPUCommandBuffer 和 GPURenderBundle 是用户录制命令的容器。GPUShaderModule 包含 着色器代码。其他资源如 GPUSampler 或 GPUBindGroup 用于配置 GPU 如何使用物理资源。

GPU 通过 GPUCommandBuffer 编码的命令，通过 管线 执行，该管线是固定功能与可编程阶段的混合体。可编程阶段运行着色器，即专为 GPU 硬件设计的特殊程序。管线的大部分状态由 GPURenderPipeline 或 GPUComputePipeline 对象定义。未包含在这些 管线对象中的状态，则在命令编码阶段通过如 beginRenderPass() 或 setBlendConstant() 之类命令设置。

2. 恶意使用考量

本节为非规范性内容。 介绍了在 Web 上暴露此 API 所带来的风险。

2.1. 安全考量

WebGPU 的安全要求与 Web 一贯的要求相同，也是不可协商的。总体原则是：在所有命令到达 GPU 之前进行严格校验，确保页面只能操作自身数据。

2.1.1. 基于 CPU 的未定义行为

WebGPU 实现会将用户发起的工作量转换为目标平台特定的 API 命令。原生 API 规定了命令的有效用法（例如见 vkCreateDescriptorSetLayout），并且通常不保证不遵守有效用法规则时的任何结果。这被称为“未定义行为”，攻击者可以利用它访问非授权内存，或迫使驱动执行任意代码。

为禁止不安全用法，WebGPU 针对任意输入都定义了允许的行为范围。实现必须校验所有用户输入，只允许有效工作负载到达驱动。本规范规定了所有错误条件及其处理语义。例如，在 copyBufferToBuffer() 的 "source" 和 "destination" 同时指定同一缓冲区，且区间相交时，GPUCommandEncoder 会生成错误，且不会执行其它操作。

更多错误处理信息参见 § 22 错误与调试。

2.1.2. 基于 GPU 的未定义行为

WebGPU 着色器 由 GPU 内部的计算单元执行。在原生 API 中，某些着色器指令可能在 GPU 上导致未定义行为。为应对这一点，WebGPU 严格定义了着色器指令集及其行为。当为 createShaderModule() 提供着色器时，WebGPU 实现必须在进行任何平台特定着色器转换或变换之前进行校验。

2.1.3. 未初始化数据

一般来说，分配新内存可能暴露系统上其它应用残留的数据。为避免此问题，WebGPU 在概念上会将所有资源初始化为零，尽管若实现检测到开发者已手动初始化内容时可跳过此步骤。这包括着色器内变量和共享工作组内存。

清除工作组内存的具体机制依平台而异。若原生 API 未提供清除机制，WebGPU 实现会先在所有调用中清空，再同步，然后继续执行开发者代码。

2.1.4. 着色器中的越界访问

着色器可以直接访问物理资源（如 "uniform" GPUBufferBinding），也可以通过纹理单元（为纹理坐标转换而设的固定功能硬件块）间接访问。WebGPU API 的校验只能保证所有着色器输入已提供且类型正确。如果未通过纹理单元访问，WebGPU API 无法保证数据访问不会越界。

为防止着色器访问非本应用 GPU 内存，WebGPU 实现可在驱动中启用“健壮缓冲区访问”模式，确保访问不会超出缓冲区界限。

或者，实现可在着色器代码中插入手动越界检查。此时，越界检查仅适用于数组索引，对结构体字段访问则无需手动检查，因为主机侧 minBindingSize 校验已覆盖。

若着色器尝试读取物理资源边界外的数据，实现可以：

1. 返回资源边界内其他位置的值

2. 返回值向量 "(0, 0, 0, X)"，其中 X 可为任意值

3. 部分丢弃绘制或调度调用

若着色器尝试写入物理资源边界外，实现可以：

1. 写入资源边界内其他位置

2. 丢弃写入操作

3. 部分丢弃绘制或调度调用

2.1.5. 无效数据

从 CPU 向 GPU 上传 浮点数数据或在 GPU 上生成浮点数时，可能会出现不对应于有效数值的二进制表示（如无穷大或 NaN）。此时 GPU 行为取决于硬件对 IEEE-754 标准的实现精度。WebGPU 保证引入无效浮点数只会影响算术运算结果，不会有其他副作用。

2.1.6. 驱动漏洞

GPU 驱动程序和其他软件一样可能存在漏洞。攻击者可能利用驱动的错误行为访问非授权数据。为降低风险，WebGPU 工作组将与 GPU 厂商协作，将 WebGPU 一致性测试套件（CTS）集成到驱动测试流程中，类似 WebGL 的做法。WebGPU 实现应对已发现的部分漏洞有兼容性处理，并对无法绕过的已知漏洞驱动禁用 WebGPU。

2.1.7. 时序攻击

2.1.7.1. 内容时间线时序

WebGPU 不会向 JavaScript 暴露在 内容时间线上由 agent 在 agent cluster 中共享的新状态。内容时间线状态如 [[mapping]] 仅在显式 内容时间线任务（如普通 JavaScript）中变化。

2.1.7.2. 设备/队列时间线时序

可写存储缓冲区和其他跨调用通信机制可能用于在队列时间线上构造高精度定时器。

可选的 "timestamp-query" 特性也为 GPU 操作提供高精度计时。为缓解安全和隐私风险，定时查询的值被对齐到较低精度：参见 current queue timestamp。特别注意：

• 设备时间线通常在多个源共享的进程中运行，因此跨源隔离（如 COOP/COEP）无法隔离设备/队列时间线定时器。

• 队列时间线工作由设备时间线发起，可能在不能像 CPU 进程（如 Meltdown 缓解）那样隔离的 GPU 硬件上执行。

• GPU 硬件通常不易受到 Spectre 类攻击，但 WebGPU 可能在软件中实现，软件实现可能运行在共享进程中，无法依赖隔离缓解。

2.1.8. Row hammer 攻击

Row hammer 是一类利用 DRAM 单元状态泄漏的攻击，可被用于 GPU。WebGPU 没有专门的防护措施，依赖于平台级解决方案，例如缩短内存刷新间隔。

2.1.9. 拒绝服务

WebGPU 应用可访问 GPU 内存和计算单元。WebGPU 实现可以限制应用可用的 GPU 内存，以保证其他应用的响应性。对于 GPU 处理时间，WebGPU 实现可以设置“看门狗”定时器，确保应用不会导致 GPU 无响应超过数秒。这些措施类似于 WebGL 所采用的策略。

2.1.10. 工作负载识别

WebGPU 提供对全局受限资源的访问，这些资源由同一台机器上不同程序（和网页）共享。应用可通过间接探测全局资源受限程度，进而推测其他网页正在进行的工作负载。这些问题与 Javascript 中的系统内存和 CPU 执行吞吐量类似。WebGPU 没有对此提供额外的缓解措施。

2.1.11. 内存资源

WebGPU 允许从机器全局内存堆（如 VRAM）进行可失败的分配。这使得应用可以通过尝试分配并观察分配失败情况，探测系统剩余可用内存（针对特定堆类型）。

GPU 内部一般有一个或多个（通常只有两个）所有运行应用共享的内存堆。当某个堆耗尽时，WebGPU 创建资源会失败。这是可观察到的，可能使恶意应用推测出其他应用使用了哪些堆，以及它们分配了多少。

2.1.12. 计算资源

如果一个站点与另一个站点同时使用 WebGPU，它可能会观察到处理某些工作的耗时增加。例如，持续向队列提交计算工作并跟踪完成时间，可能会发现其他任务也开始使用 GPU。

GPU 有多个可独立测试的部分，如算术单元、纹理采样单元、原子单元等。恶意应用可以检测某些单元的负载，并试图分析压力模式来猜测其他应用的工作负载。这与 Javascript 的 CPU 执行现实类似。

2.1.13. 能力滥用

恶意站点可能滥用 WebGPU 所暴露的能力，运行对用户或其体验无益、仅对站点有利的计算。例如隐蔽挖矿、密码破解或彩虹表计算。

无法针对这类 API 使用方式加以防范，因为浏览器无法区分有效负载和滥用负载。这是 Web 上所有通用计算能力（如 JavaScript、WebAssembly 或 WebGL）面临的普遍问题。WebGPU 只是让某些负载的实现更容易，或比 WebGL 更高效。

为缓解此类滥用，浏览器可对后台标签页的操作进行限速，可警告标签页资源占用过高，并可限制哪些上下文允许使用 WebGPU。

用户代理可基于启发式方法向用户发出高功耗警告，尤其在检测到潜在恶意使用时。如果实现此类警告，应将 WebGPU 使用纳入与 JavaScript、WebAssembly、WebGL 等相同的启发式中。

2.2. 隐私考量

WebGPU 的隐私考量与 WebGL 类似。GPU API 十分复杂，必须出于必要暴露设备能力的各个方面，以便开发者能够有效利用这些能力。一般的缓解方法是对潜在可识别信息进行归一化或分桶，并在可能的情况下强制行为一致。

用户代理不得暴露超过 32 种可区分配置或分桶。

2.2.1. 机器特有功能与限制

WebGPU 可以揭示底层 GPU 架构和设备结构的许多细节，包括可用的物理适配器、GPU 和 CPU 资源的多项限制（如最大纹理尺寸），以及可用的任何可选硬件专属能力。

用户代理没有义务暴露真实硬件限制，可以完全控制机器细节的暴露程度。减少指纹识别的一种策略是将所有目标平台归为少数几个分桶。总体来说，暴露硬件限制的隐私影响与 WebGL 相同。

默认限制值也被故意设置得足够高，以便大多数应用无需请求更高限制即可运行。所有 API 的使用都按请求限制进行校验，因此不会因意外而向用户暴露实际硬件能力。

2.2.2. 机器特有产物

和 WebGL 一样，可以观察到一些机器特有的光栅化/精度差异和性能差异。这涉及光栅化覆盖及模式、着色器阶段间 varyings 的插值精度、计算单元调度，以及更多执行相关特性。

通常，同一厂商的大多数或全部设备的光栅化和精度指纹是相同的。性能差异难以完全规避，但信号强度较低（如 JS 执行性能）。

对隐私要求高的应用和用户代理应采用软件实现以消除这类产物。

2.2.3. 机器特有性能

通过测量 GPU 上特定操作的性能也是区分用户的一个因素。即便计时精度较低，重复执行某一操作也能体现用户机器在特定负载下的快慢。这是常见的识别向量（WebGL 和 Javascript 皆有），但信号较低且难以完全规避。

WebGPU 计算管线让开发者可绕过固定功能硬件直接访问 GPU，这带来了独特设备指纹识别的额外风险。用户代理可通过将逻辑 GPU 调用与实际计算单元解耦来降低此风险。

2.2.4. 用户代理状态

本规范未为源定义任何额外用户代理状态。但预期用户代理会对 GPUShaderModule、 GPURenderPipeline、 GPUComputePipeline 等高开销编译结果进行缓存。这些缓存有助于提升 WebGPU 应用首次访问后的加载速度。

对规范而言，这些缓存与极快编译无异，但应用可以轻易测量 createComputePipelineAsync() 的耗时。这可能导致跨源信息泄露（如“用户是否访问过包含特定着色器的站点”），因此用户代理应遵循 存储分区最佳实践。

系统的 GPU 驱动也可能有自己的着色器和管线编译缓存。用户代理可在可能的情况下禁用该缓存，或为着色器添加分区数据，使驱动将其视为不同对象。

2.2.5. 驱动漏洞

除安全考量中提到的问题外，驱动漏洞还可能导致行为差异，成为区分用户的手段。此处可采用安全考量中提及的缓解措施，如与 GPU 厂商协作，在用户代理中实现已知问题的兼容处理。

2.2.6. 适配器标识符

WebGL 实践表明，开发者有合理需求识别其代码运行的 GPU，以便创建和维护健壮的 GPU 内容。例如识别有已知驱动漏洞的适配器，以便规避或避免在某些硬件上使用表现不佳的特性。

但暴露适配器标识符自然会增加指纹识别信息量，因此有必要限制适配器识别的精度。

可以采取多项缓解措施以平衡内容健壮性与隐私保护。首先，用户代理可通过主动识别和规避已知驱动问题，减少开发者负担，就像浏览器开始用 GPU 以来所做的一样。

默认暴露适配器标识符时应尽量宽泛，只要有用即可。例如，可仅标识适配器厂商及架构而非具体型号。有时也可报告实际适配器的合理代理的标识符。

在需要完整详细适配器信息（如提交 bug 报告）时，可请用户同意向页面披露更多硬件信息。

最后，若用户代理认为合适（如增强隐私模式下），可完全不报告适配器标识符。

3. 基础

3.1. 约定

3.1.1. 语法速记

本规范中使用了如下语法速记：

.（点）语法，常见于编程语言。

短语“Foo.Bar”表示“值（或接口）Foo 的 Bar 成员”。如果 Foo 是 有序映射（ordered map）且 Bar 在 Foo 中未存在，则返回 undefined。

短语“Foo.Bar 被提供”表示“Bar 成员存在于 映射值 Foo 中”。

?.（可选链）语法，借鉴自 JavaScript。

短语“Foo?.Bar”表示：“如果 Foo 为 null 或 undefined 或 Bar 在 Foo 中未存在，则为 undefined；否则为 Foo.Bar”。

例如，若 buffer 是一个 GPUBuffer，则 buffer?.\[[device]].\[[adapter]] 表示：“如果 buffer 为 null 或 undefined，则为 undefined；否则为 buffer 的 \[[device]] 内部槽的 \[[adapter]] 内部槽。”

??（空值合并）语法，借鉴自 JavaScript。

短语“x ?? y”表示：“如果 x 不为 null 或 undefined，则为 x，否则为 y。”

槽支持属性（slot-backed attribute）

由同名内部槽支持的 WebIDL 属性。其可变性视规范而定。

3.1.2. WebGPU 对象

WebGPU 对象由WebGPU 接口和内部对象组成。

WebGPU 接口定义了WebGPU 对象的公共接口和状态。它可在其创建的内容时间线上使用，此时它是 JavaScript 暴露的 WebIDL 接口。

任何包含 GPUObjectBase 的接口都是 WebGPU 接口。

内部对象跟踪WebGPU 对象在设备时间线上的状态。所有对内部对象可变状态的读写都发生在单一有序的设备时间线步骤中。

在WebGPU 对象上可定义以下特殊属性类型：

不可变属性（immutable property）

在对象初始化时设定的只读槽，可从任意时间线访问。

注：由于该槽不可变，实现可在多个时间线中保有副本，按需分配。不可变属性这样定义是为避免在规范中描述多个副本。

若命名为 [[带括号]]，则是内部槽。
若命名为 不带括号，则是 WebGPU 接口的 readonly 槽支持属性。

内容时间线属性（content timeline property）

仅可从对象创建的内容时间线访问的属性。

若命名为 [[带括号]]，则是内部槽。
若命名为 不带括号，则是 WebGPU 接口的槽支持属性。

设备时间线属性（device timeline property）

跟踪内部对象状态，仅可从对象创建的设备时间线访问。设备时间线属性可为可变的。

设备时间线属性命名为 [[带括号]]，为内部槽。

队列时间线属性（queue timeline property）

跟踪内部对象状态，仅可从对象创建的队列时间线访问。队列时间线属性可为可变的。

队列时间线属性命名为 [[带括号]]，为内部槽。

interface mixin GPUObjectBase {
    attribute USVString label;
};

GPUObjectBase 具有如下不可变属性：

[[device]]，类型为device，只读

拥有该内部对象的设备。

对此对象内容的操作断言其运行在设备时间线上，并且设备是有效的。

GPUObjectBase 具有如下内容时间线属性：

label，类型为 USVString

开发者提供的标签，由实现自定义使用。可由浏览器、操作系统或其他工具用于帮助开发者识别底层内部对象。示例包括在 GPUError 消息、控制台警告、浏览器开发者工具和平台调试工具中显示标签。

注：

实现应当利用标签提升错误消息的可读性，用于标识 WebGPU 对象。

但这不必是标识对象的唯一方式：实现还应结合其他可用信息，尤其是在没有标签时。例如：

• 打印 GPUTexture 视图时，显示其父对象的标签。

• 打印 GPURenderPassEncoder 或 GPUComputePassEncoder 时，显示其父 GPUCommandEncoder 的标签。

• 打印 GPUCommandBuffer 时，显示其源 GPUCommandEncoder 的标签。

• 打印 GPURenderBundle 时，显示其源 GPURenderBundleEncoder 的标签。

注：

label 是 GPUObjectBase 的属性。两个 GPUObjectBase “包装器”对象即使引用同一底层对象，其 label 状态也是完全独立的（如由 getBindGroupLayout() 返回时）。除非通过 JavaScript 设置，否则 label 属性不会变更。

这意味着一个底层对象可关联多个标签。本规范未定义标签如何传递到 设备时间线。标签的使用完全由实现自定义：错误消息可显示最近设置的标签、所有已知标签，或完全不显示标签。

属性类型为 USVString，因为某些用户代理可能将其传递给底层原生 API 的调试工具。

GPUObjectBase 具有如下设备时间线属性：

[[valid]]，类型为boolean，初始值为 true

若为 true，表示该内部对象可用。

3.1.3. 对象描述符

对象描述符用于保存创建对象所需的信息，通常通过 GPUDevice 的 create* 方法之一进行对象创建。

dictionary GPUObjectDescriptorBase {
    USVString label = "";
};

GPUObjectDescriptorBase 具有如下成员：

label，类型为 USVString，默认值为 ""

GPUObjectBase.label 的初始值。

3.2. 异步性

3.2.1. 无效的内部对象与传染性无效

WebGPU 中的对象创建操作不会返回 Promise，但其本质上是异步的。返回的对象引用的是在设备时间线上被操作的内部对象。多数在设备时间线上发生的错误不会通过异常或拒绝抛出，而是通过关联设备上生成的 GPUError 进行传递。

内部对象可以是有效或无效。无效对象不会在之后变为有效，但部分有效对象可以被使无效（invalidated）。

如果对象无法被创建，则其从创建开始即为无效。例如，对象描述符未能描述出有效对象，或者没有足够内存分配资源时会发生此情况。如果从另一个无效对象创建对象（例如对无效的 GPUTexture 调用 createView()），也会如此。这类情况称为传染性无效。

大多数类型的内部对象在创建后不会变为无效，但仍可能变得不可用，例如其所属设备丢失或destroyed，或对象处于特殊内部状态（如缓冲区状态“destroyed”）。

部分类型的内部对象在创建后可以变为无效；具体包括设备、适配器、GPUCommandBuffer，以及命令/通道/捆绑编码器。

3.2.2. Promise 顺序

WebGPU 中有若干操作会返回 promise。

• GPU.requestAdapter()

• GPUAdapter.requestDevice()

• GPUDevice.createComputePipelineAsync()

• GPUDevice.createRenderPipelineAsync()

• GPUShaderModule.getCompilationInfo()

• GPUQueue.onSubmittedWorkDone()

• GPUBuffer.mapAsync()

• GPUDevice.lost

• GPUDevice.popErrorScope()

WebGPU 不保证这些 promise 的 settle（resolve 或 reject）顺序，除非有如下例外：

• 对于某个 GPUQueue q，如果 p1 = q.onSubmittedWorkDone() 先于 p2 = q.onSubmittedWorkDone() 被调用，则 p1 必须早于 p2 settle。
• 对于某个 GPUQueue q 和同一 GPUDevice 上的 GPUBuffer b，如果 p1 = b.mapAsync() 先于 p2 = q.onSubmittedWorkDone() 被调用，则 p1 必须早于 p2 settle。

应用不得依赖于其它 promise 的 settle 顺序。

3.3. 坐标系统

渲染操作使用如下坐标系统：

• 归一化设备坐标（NDC）有三维：

  • -1.0 ≤ x ≤ 1.0

  • -1.0 ≤ y ≤ 1.0

  • 0.0 ≤ z ≤ 1.0

  • 左下角为 (-1.0, -1.0, z)。

  注： z = 0 还是 z = 1 作为近平面由应用决定。上图以 z = 0 为近平面，但实际行为由着色器使用的投影矩阵、depthClearValue 和 depthCompare 函数共同决定。

• 裁剪空间坐标有四维：(x, y, z, w)

  • 裁剪空间坐标用于顶点的裁剪位置（即顶点着色器的position输出），以及裁剪体积。

  • 归一化设备坐标和裁剪空间坐标的关系如下：若点 p = (p.x, p.y, p.z, p.w) 在裁剪体积内，则其归一化设备坐标为 (p.x ÷ p.w, p.y ÷ p.w, p.z ÷ p.w)。

• 帧缓冲坐标用于寻址帧缓冲内的像素。

  • 为二维坐标。

  • 每个像素在 x、y 方向上长度为 1。

  • 左上角为 (0.0, 0.0)。

  • x 向右递增。

  • y 向下递增。

  • 参见 § 17 渲染通道 和 § 23.2.5 光栅化。

• 视口坐标结合了 x、y 方向的帧缓冲坐标与 z 方向的深度。

  • 通常 0.0 ≤ z ≤ 1.0，但可通过 [[viewport]].minDepth 和 maxDepth，以及 setViewport() 修改。

• 片元坐标与视口坐标一致。

• 纹理坐标，在 2D 中有时称为“UV 坐标”，用于采样纹理，其分量数量与 texture dimension 匹配。

  • 0 ≤ u ≤ 1.0

  • 0 ≤ v ≤ 1.0

  • 0 ≤ w ≤ 1.0

  • (0.0, 0.0, 0.0) 位于纹理存储顺序的首个 texel。

  • (1.0, 1.0, 1.0) 位于纹理存储顺序的最后一个 texel。

• 窗口坐标，或称显示坐标，与帧缓冲坐标一致，用于与外部显示或类似接口交互时。

注：WebGPU 的坐标系统与 DirectX 图形管线中的坐标系统一致。

3.4. 编程模型

3.4.1. 时间线

WebGPU 的行为以“时间线”为单位进行描述。每个操作（以算法定义）都发生在某个时间线上。时间线明确了操作的顺序，以及哪些状态可被哪些操作访问。

注： 这种“时间线”模型反映了浏览器引擎多进程模型（如“内容进程”和“GPU 进程”）的约束，也体现了许多实现中 GPU 作为独立执行单元的现实。实现 WebGPU 并不要求时间线并行执行，因此不要求多进程，甚至不要求多线程。（但如 获取上下文图像内容副本 这类需同步阻塞其它时间线完成的情况，仍需支持并发。）

内容时间线

与 Web 脚本的执行相关。包括本规范描述的所有方法调用。

要从 GPUDevice device 的操作向内容时间线派发步骤，可通过 为 GPUDevice 排队全局任务。

设备时间线

与用户代理发起的 GPU 设备操作相关。包括适配器、设备、GPU 资源与状态对象的创建，这些操作从用户代理控制 GPU 的部分看通常是同步的，但可在独立的操作系统进程中运行。

队列时间线

与 GPU 计算单元上操作的执行相关。包括实际在 GPU 上运行的绘制、拷贝和计算任务。

时间线无关

可以与上述任意时间线关联。

如仅操作不可变属性或调用步骤传入的参数，则可派发到任意时间线。

本规范中，当返回值依赖于非内容时间线上发生的工作时，采用异步操作。API 以 promise 或事件的形式表现异步操作。

3.4.2. 内存模型

本节为非规范性内容。

一旦应用初始化过程中获取到 GPUDevice，我们可以将 WebGPU 平台描述为包含以下各层：

1. 实现本规范的用户代理。

2. 为该设备提供底层原生 API 驱动的操作系统。

3. 实际的 CPU 和 GPU 硬件。

WebGPU 平台的每一层可能有不同类型的内存，用户代理在实现规范时需要加以考虑：

• 脚本拥有的内存，如脚本创建的 ArrayBuffer，通常无法被 GPU 驱动直接访问。

• 用户代理可能有不同进程分别负责内容运行和与 GPU 驱动的通信。在这种情况下，会使用进程间共享内存来传递数据。

• 独立显卡有自己的高带宽内存，而集成显卡通常与系统共享内存。

大多数物理资源分配在适合 GPU 计算或渲染的内存类型中。当用户需要向 GPU 提供新数据时，数据可能首先需要跨越进程边界，抵达与 GPU 驱动通信的用户代理部分。随后可能需要使数据对驱动可见，有时这需要拷贝到驱动分配的中转（staging）内存。最后，数据还可能需要传输到独立 GPU 内存中，并在内部转换为最适合 GPU 运算的布局。

所有这些转换都由用户代理的 WebGPU 实现负责完成。

注：上述示例描述了最坏情况，实际实现中可能无需跨越进程边界，或者能直接向用户暴露由驱动管理的内存（例如通过 ArrayBuffer），从而避免数据拷贝。

3.4.3. 资源用法

物理资源可以被GPU 命令以内部用法使用：

输入（input）

为 draw 或 dispatch 调用提供输入数据的缓冲区。会保留内容。由缓冲区 INDEX、VERTEX 或 INDIRECT 允许。

常量（constant）

从着色器视角为常量的资源绑定。会保留内容。由缓冲区 UNIFORM 或纹理 TEXTURE_BINDING 允许。

存储（storage）

读/写存储资源绑定。由缓冲区 STORAGE 或纹理 STORAGE_BINDING 允许。

存储只读（storage-read）

只读存储资源绑定。会保留内容。由缓冲区 STORAGE 或纹理 STORAGE_BINDING 允许。

附件（attachment）

在渲染通道中作为读/写输出附件或仅写解决目标的纹理。由纹理 RENDER_ATTACHMENT 允许。

附件只读（attachment-read）

在渲染通道中作为只读附件的纹理。会保留内容。由纹理 RENDER_ATTACHMENT 允许。

子资源（subresource）指整个缓冲区或纹理子资源。

强制要求用法仅可组合为兼容用法列表，使得 API 能够限制内存操作中的数据竞争出现时机。该属性使基于 WebGPU 的应用更容易无修改地在不同平台上运行。

3.4.4. 同步机制

用法范围（usage scope）是一个从有序映射，其键为子资源，值为 列表<内部用法>。每个用法范围覆盖一组可能并发执行的操作，因此在该范围内对子资源的使用只能是兼容用法列表。

用法范围在编码期间被构建和校验：

• 在 dispatchWorkgroups() 中

• 在 dispatchWorkgroupsIndirect() 中

• 在 GPURenderPassEncoder.end() 时

• 在 GPURenderBundleEncoder.finish() 时

用法范围如下：

• 在计算通道中，每个 dispatch 命令（dispatchWorkgroups() 或 dispatchWorkgroupsIndirect()）即为一个用法范围。

  若某子资源在当前 dispatch 调用中可能被访问，则视为在该用法范围内被使用，包括：

  • 当前 GPUComputePipeline 的 [[layout]] 所用绑定组（bind group）槽引用的全部子资源

  • dispatch 调用直接使用的缓冲区（如间接缓冲区）

  注： 诸如 setBindGroup() 等状态设定命令本身不会直接将绑定资源计入用法范围，它们只改变 dispatch 命令检查的状态。

• 一次渲染通道（render pass）即为一个用法范围。

  若某子资源在任何命令中被引用（包括状态设定命令，与计算通道不同），则视为在该用法范围内被使用，包括：

  • 通过 setVertexBuffer() 设置的缓冲区

  • 通过 setIndexBuffer() 设置的缓冲区

  • 通过 setBindGroup() 设置的绑定组引用的所有子资源

  • draw 调用直接使用的缓冲区（如间接缓冲区）

注：拷贝命令为独立操作，不使用用法范围进行校验，其自身实现了防止自竞争的校验。

3.5. 核心内部对象

3.5.1. 适配器（Adapters）

适配器（adapter）标识系统上的 WebGPU 实现：既包括底层平台上的计算/渲染功能实例，也包括浏览器在该功能之上实现 WebGPU 的实例。

适配器通过 GPUAdapter 暴露。

适配器并不唯一代表底层实现：多次调用 requestAdapter() 每次都会返回不同的适配器对象。

每个 适配器对象只能用于创建一个 设备（device）：一旦成功调用 requestDevice()，该适配器的 [[state]] 变为 "consumed"。此外，适配器对象可能在任何时刻过期（expire）。

注： 这样保证应用在创建设备时能使用最新的系统状态来选择适配器。也让多种场景（如首次初始化、因适配器移除重初始化、因 GPUDevice.destroy() 测试重初始化等）行为保持一致，提高健壮性。

若适配器以显著性能换取更广兼容性、更可预测行为或更好的隐私，则可视为回退适配器（fallback adapter）。不是每个系统都必须提供回退适配器。

适配器具有如下不可变属性：

[[features]]，类型为 有序集合<GPUFeatureName>，只读

可用于在此适配器上创建设备的特性。

[[limits]]，类型为支持的限制，只读

可用于在此适配器上创建设备的最佳限制。每个限制值必须与支持的限制中该项的默认值相同或更好（better）。

[[fallback]]，类型为 boolean，只读

为 true 时，表示该适配器是回退适配器。

[[xrCompatible]]，类型为 boolean

为 true 时，表示该适配器请求了与 WebXR 会话 兼容。

适配器具有如下设备时间线属性：

[[state]]，初始值为 "valid"

"valid"

适配器可用于创建设备。

"consumed"

适配器已被用于创建设备，不能再次使用。

"expired"

适配器因其他原因已过期。

3.5.2. 设备（Devices）

设备（device）是适配器的逻辑实例，通过它可以创建内部对象。

设备通过 GPUDevice 暴露。

一个设备独占其上创建的所有内部对象：当该设备变为无效（丢失或 销毁时），它本身和所有直接（如 createTexture()）或间接（如 createView()）创建的对象都会隐式变为不可用。

设备具有如下不可变属性：

[[adapter]]，类型为 适配器，只读

创建该设备的适配器。

[[features]]，类型为 有序集合<GPUFeatureName>，只读

可在该设备上使用的特性，在创建时确定。即使底层适配器支持更多特性，也不能使用额外特性。

[[limits]]，类型为支持的限制，只读

可在该设备上使用的限制，在创建时确定。即使底层适配器支持更高限制，也不能使用更高值（better）。

设备具有如下内容时间线属性：

[[content device]]，类型为 GPUDevice，只读

该设备关联的内容时间线 GPUDevice 接口实例。

每当用户代理需要撤销设备访问权限时，会在该设备的设备时间线上调用 丢失设备(device, "unknown")，该操作可能优先于当前队列中的其他操作。

如果某操作失败且副作用可能导致设备上的对象状态可见变化或内部实现/驱动状态损坏，应当丢失该设备以防止这些变化被观察到。

注： 非应用主动发起的设备丢失（通过 destroy()）时，用户代理应无条件向开发者发出警告，即使 lost promise 已被处理。此类场景应极其罕见，并且该信号对开发者至关重要，因为大多数 WebGPU API 会尽量表现为“一切正常”以避免中断运行时流程：不会抛出校验错误，大多数 promise 正常 resolve 等。

3.6. 可选功能

WebGPU 适配器和设备具备功能，这些功能描述了WebGPU在不同实现之间的差异， 通常是由于硬件或系统软件的限制。 一项功能可以是特性或限制。

用户代理不得透露超过32个可区分的配置或分组。

适配器的功能必须符合§ 4.2.1 适配器功能保证。

仅支持的功能可以在requestDevice()中请求； 请求不支持的功能会导致失败。

设备的功能在"一个新的设备"中确定， 通过从适配器的默认值（无特性和默认的支持的限制）开始， 并根据在requestDevice()中请求的功能添加功能。 这些功能无论适配器的功能如何均会被强制执行。

有关隐私方面的考虑，请参阅§ 2.2.1 机器特定的功能和限制。

3.6.1. 特性

特性是一组可选的 WebGPU 功能，并非所有实现都支持这些功能，通常是由于硬件或系统软件的限制所致。

所有特性都是可选的，但适配器会对其可用性做出一定保证 （参见§ 4.2.1 适配器功能保证）。

设备仅支持在创建时确定的那组特性（参见§ 3.6 可选功能）。 API 调用将根据这些特性（而非适配器的特性）进行校验：

• 以新方式使用已有的 API 接口通常会导致校验错误。

• 有几种类型的可选 API 接口：

  • 使用新方法或枚举值时总是抛出 TypeError 异常。

  • 使用带有（正确类型）非默认值的新字典成员通常会导致校验错误。

  • 使用新的 WGSL enable 指令时，总是会在 createShaderModule() 的过程中产生校验错误。

每个特性所启用的功能描述可参见特性索引。

注意： 启用特性未必是理想选择，可能会影响性能。 因此，为了提升不同设备和实现间的可移植性，应用程序通常只应请求实际需要的特性。

3.6.2. 限制

每个限制都是对设备上 WebGPU 使用的数值限制。

注意： 设置“更好”的限制未必是理想选择，因为这样可能会对性能产生影响。 因此，为了提升不同设备和实现之间的可移植性，应用程序通常只有在确实需要时才请求比默认值更好的限制。

每个限制都有一个默认值。

适配器始终保证支持默认值或更好的 （参见§ 4.2.1 适配器功能保证）。

设备仅支持在创建时确定的那组限制（参见§ 3.6 可选功能）。 API 调用根据这些限制进行校验（而非适配器的限制）， 不允许“更好”或更差。

对于任何给定的限制，一些值是更好的。 更好的限制值始终放宽校验，从而使更多程序变得有效。 对于每个限制类别，“更好”的定义如下。

不同的限制具有不同的限制类别：

最大值

限制在传递给 API 的某些值上强制设置一个最大值。

更高的值是更好的。

仅能设置为 ≥ 默认值 的值。 较低的值会被限制为默认值。

对齐

限制在传递给 API 的某些值上强制设置一个最小对齐；即该值必须是限制的倍数。

较低的值是更好的。

仅能设置为 ≤ 默认值 的2的幂次。 非2的幂次值是无效的。 更高的2的幂次值会被限制为默认值。

支持的限制 对象具有 WebGPU 定义的每个限制的值：

3.6.2.1. GPUSupportedLimits

GPUSupportedLimits 用于暴露适配器或设备的支持的限制。 参见 GPUAdapter.limits 和 GPUDevice.limits。

[Exposed=(Window, Worker), SecureContext]
interface GPUSupportedLimits {
    readonly attribute unsigned long maxTextureDimension1D;
    readonly attribute unsigned long maxTextureDimension2D;
    readonly attribute unsigned long maxTextureDimension3D;
    readonly attribute unsigned long maxTextureArrayLayers;
    readonly attribute unsigned long maxBindGroups;
    readonly attribute unsigned long maxBindGroupsPlusVertexBuffers;
    readonly attribute unsigned long maxBindingsPerBindGroup;
    readonly attribute unsigned long maxDynamicUniformBuffersPerPipelineLayout;
    readonly attribute unsigned long maxDynamicStorageBuffersPerPipelineLayout;
    readonly attribute unsigned long maxSampledTexturesPerShaderStage;
    readonly attribute unsigned long maxSamplersPerShaderStage;
    readonly attribute unsigned long maxStorageBuffersPerShaderStage;
    readonly attribute unsigned long maxStorageTexturesPerShaderStage;
    readonly attribute unsigned long maxUniformBuffersPerShaderStage;
    readonly attribute unsigned long long maxUniformBufferBindingSize;
    readonly attribute unsigned long long maxStorageBufferBindingSize;
    readonly attribute unsigned long minUniformBufferOffsetAlignment;
    readonly attribute unsigned long minStorageBufferOffsetAlignment;
    readonly attribute unsigned long maxVertexBuffers;
    readonly attribute unsigned long long maxBufferSize;
    readonly attribute unsigned long maxVertexAttributes;
    readonly attribute unsigned long maxVertexBufferArrayStride;
    readonly attribute unsigned long maxInterStageShaderVariables;
    readonly attribute unsigned long maxColorAttachments;
    readonly attribute unsigned long maxColorAttachmentBytesPerSample;
    readonly attribute unsigned long maxComputeWorkgroupStorageSize;
    readonly attribute unsigned long maxComputeInvocationsPerWorkgroup;
    readonly attribute unsigned long maxComputeWorkgroupSizeX;
    readonly attribute unsigned long maxComputeWorkgroupSizeY;
    readonly attribute unsigned long maxComputeWorkgroupSizeZ;
    readonly attribute unsigned long maxComputeWorkgroupsPerDimension;
};

3.6.2.2. GPUSupportedFeatures

GPUSupportedFeatures 是一个 类似集合（setlike） 接口。它的 集合条目 是适配器或设备支持的 GPUFeatureName 值。 它只能包含 GPUFeatureName 枚举中的字符串。

[Exposed=(Window, Worker), SecureContext]
interface GPUSupportedFeatures {
    readonly setlike<DOMString>;
};

if (adapter.features.has('unknown-feature')) {
    // 使用 unknown-feature
} else {
    console.warn('unknown-feature is not supported by this adapter.');
}

3.6.2.3. WGSLLanguageFeatures

WGSLLanguageFeatures 是 navigator.gpu.wgslLanguageFeatures 的 类似集合（setlike） 接口。 它的 集合条目 是该实现支持的 WGSL 语言扩展 的字符串名称（不论适配器或设备）。

[Exposed=(Window, Worker), SecureContext]
interface WGSLLanguageFeatures {
    readonly setlike<DOMString>;
};

3.6.2.4. GPUAdapterInfo

GPUAdapterInfo 用于暴露关于适配器的各种标识信息。

GPUAdapterInfo 的成员不保证有任何特定值；如果没有提供值，该属性将返回空字符串 ""。是否揭示这些值由用户代理自行决定，并且某些设备上可能所有值都为空。因此，应用必须能够处理任何可能的 GPUAdapterInfo 值，包括这些值缺失的情况。

适配器的 GPUAdapterInfo 可通过 GPUAdapter.info 和 GPUDevice.adapterInfo 获取。 这些信息是不可变的：对于某个适配器，每次访问同一个 GPUAdapterInfo 属性都会返回相同的值。

注意: 虽然 GPUAdapterInfo 属性一旦访问就不可变， 但实现可以在首次访问前延后决定每个属性的内容。

注意: 即便其他 GPUAdapter 实例代表同一物理适配器， 它们在 GPUAdapterInfo 中揭示的值也可能不同。 但除非有特殊事件提升了页面可访问的标识信息（本规范未定义此类事件），否则它们应当揭示相同的值。

关于隐私考虑，请参见 § 2.2.6 适配器标识符。

[Exposed=(Window, Worker), SecureContext]
interface GPUAdapterInfo {
    readonly attribute DOMString vendor;
    readonly attribute DOMString architecture;
    readonly attribute DOMString device;
    readonly attribute DOMString description;
    readonly attribute unsigned long subgroupMinSize;
    readonly attribute unsigned long subgroupMaxSize;
    readonly attribute boolean isFallbackAdapter;
};

GPUAdapterInfo 拥有如下属性：

vendor， 类型为 DOMString，只读

适配器的厂商名称（如有）。否则为空字符串。

architecture， 类型为 DOMString，只读

适配器所属 GPU 家族或类别的名称（如有）。否则为空字符串。

device， 类型为 DOMString，只读

适配器的厂商自定义标识符（如有）。否则为空字符串。

注意: 这是代表适配器类型的值。例如，可能是 PCI 设备ID。它不会像序列号一样唯一标识某块硬件。

description， 类型为 DOMString，只读

驱动报告的关于适配器的人类可读字符串描述（如有）。否则为空字符串。

注意: description 没有采用任何格式化，不建议尝试解析此值。根据 GPUAdapterInfo 改变行为的应用（如为已知驱动问题应用修正），应尽量依赖其他字段。

subgroupMinSize， 类型为 unsigned long，只读

如果支持 "subgroups" 特性，则为适配器支持的最小 子组大小。

subgroupMaxSize， 类型为 unsigned long，只读

如果支持 "subgroups" 特性，则为适配器支持的最大 子组大小。

isFallbackAdapter， 类型为 boolean，只读

适配器是否为回退适配器。

3.7. 扩展文档

“扩展文档”是描述新功能的附加文档，这些功能是非规范性的，不属于 WebGPU/WGSL 规范的一部分。 它们描述了在这些规范基础上构建的功能，通常包含一个或多个新的 API 特性标志和/或 WGSL enable 指令，或与其他草案 Web 规范的交互。

WebGPU 实现不得暴露扩展功能；这样做属于规范违规。 新功能只有在被集成到 WebGPU 规范（本文档）和/或 WGSL 规范后，才成为 WebGPU 标准的一部分。

3.8. 源限制（Origin Restrictions）

WebGPU 允许访问存储在图片、视频和画布中的图像数据。 出于安全原因，对跨域媒体的使用施加了限制，因为着色器可以被用来间接推测已上传到 GPU 的纹理内容。

WebGPU 不允许上传 不是 origin-clean 的图片源。

这也意味着，使用 WebGPU 渲染的 canvas 的 origin-clean 标志永远不会被置为 false。

关于为图片和视频元素发起 CORS 请求的更多信息，请参考：

• HTML § 2.5.4 CORS settings attributes

• HTML § 4.8.3 The img element img

• HTML § 4.8.11 Media elements HTMLMediaElement

3.9. 任务源（Task Sources）

3.9.1. WebGPU 任务源

WebGPU 定义了一个新的 任务源，称为 WebGPU 任务源。 它用于 uncapturederror 事件和 GPUDevice.lost。

3.9.2. 自动过期任务源

WebGPU 定义了一个新的 任务源，称为 自动过期任务源。 它用于自动、定时地销毁某些对象：

• 由 getCurrentTexture() 返回的 GPUTexture

• 由 HTMLVideoElement 创建的 GPUExternalTexture

来自 自动过期任务源 的任务应当以高优先级处理；尤其是排入队列后，应当在用户自定义（JavaScript）任务之前执行。

3.10. 色彩空间与编码

WebGPU 不提供色彩管理。WebGPU 内的所有值（如纹理元素）都是原始数值，不是经过色彩管理的色值。

WebGPU 确实与色彩管理的输出（通过 GPUCanvasConfiguration）和输入（通过 copyExternalImageToTexture() 及 importExternalTexture()）对接。因此，必须在 WebGPU 数值和外部色值之间进行色彩转换。每个接口点都会本地定义一种编码（色彩空间、传递函数和 alpha 预乘），用于解释 WebGPU 的数值。

WebGPU 允许 PredefinedColorSpace 枚举中的所有色彩空间。注意，每个色彩空间都定义了扩展范围（详见 CSS 相关定义），可以表示其空间外的颜色值（包括色度和明度）。

超出色域的预乘 RGBA 值指 R/G/B 其中任意一个通道的值大于 alpha 通道的值。例如，预乘 sRGB RGBA 值 [1.0, 0, 0, 0.5] 表示原始（未预乘）颜色 [2, 0, 0] 且 alpha 为 50%，在 CSS 中可写作 rgb(srgb 2 0 0 / 50%)。和所有 sRGB 色域外的颜色一样，这在扩展色彩空间中是有定义的点（alpha 为 0 时除外，此时没有颜色）。但若此类值输出到可见画布，结果未定义（见 GPUCanvasAlphaMode "premultiplied"）。

3.10.1. 色彩空间转换

颜色在空间间转换时，需根据上述定义将其在一个空间中的表示转换为另一个空间中的表示。

如果源值少于4个 RGBA 通道，则缺失的绿/蓝/alpha 通道分别设为0, 0, 1，然后再做色彩空间/编码和 alpha 预乘转换。转换后，如果目标需要的通道数少于4，则忽略多余通道。

注意: 灰度图像通常在其色彩空间下表现为 RGB 值 (V, V, V) 或 RGBA 值 (V, V, V, A)。

颜色在转换时不会被有损截断：若源色值本就在目标色彩空间的色域外，转换后会产生超出[0, 1]范围的值。例如，sRGB 目标下，如果源为 rgba16float，色彩空间更广（如 Display-P3），或为预乘且包含超出色域的预乘值，都可能出现上述情况。

同样，如果源值有高位深（如 16 位/分量 PNG）或扩展范围（如 float16 存储的 canvas），这些颜色会在转换过程中保留，且中间计算的精度不少于源。

3.10.2. 色彩空间转换省略

若源与目标的色彩空间/编码一致，则无需转换。通常，若某一步是恒等函数（无操作），实现应当出于性能考虑省略该步骤。

为获得最佳性能，应用应当设置色彩空间和编码选项，以最小化整个流程中所需的转换次数。 对于各种 GPUCopyExternalImageSourceInfo 图像源：

• ImageBitmap：

  • 预乘由 premultiplyAlpha 控制。

  • 色彩空间由 colorSpaceConversion 控制。

• 2d 画布：

  • 始终为预乘。

  • 色彩空间由 colorSpace 创建上下文时的属性控制。

• WebGL 画布：

  • 预乘由 WebGLContextAttributes 中的 premultipliedAlpha 选项控制。

  • 色彩空间由 WebGLRenderingContextBase 的 drawingBufferColorSpace 状态控制。

注意: 在依赖这些特性前，请检查浏览器的实现支持。

3.11. JavaScript 到 WGSL 的数值转换

WebGPU API 的多个部分（可覆盖管线 constants 和渲染通道清除值）会接收来自 WebIDL（double 或 float）的数值，并将其转换为 WGSL 值（bool、i32、u32、f32、f16）。

4. Initialization

4.1. navigator.gpu

A GPU object is available in the Window and WorkerGlobalScope contexts through the Navigator and WorkerNavigator interfaces respectively and is exposed via navigator.gpu:

interface mixin NavigatorGPU {
    [SameObject, SecureContext] readonly attribute GPU gpu;
};
Navigator includes NavigatorGPU;
WorkerNavigator includes NavigatorGPU;

NavigatorGPU has the following attributes:

gpu, of type GPU, readonly

A global singleton providing top-level entry points like requestAdapter().

4.2. GPU

GPU is the entry point to WebGPU.

[Exposed=(Window, Worker), SecureContext]
interface GPU {
    Promise<GPUAdapter?> requestAdapter(optional GPURequestAdapterOptions options = {});
    GPUTextureFormat getPreferredCanvasFormat();
    [SameObject] readonly attribute WGSLLanguageFeatures wgslLanguageFeatures;
};

GPU has the following methods:

requestAdapter(options)

Requests an adapter from the user agent. The user agent chooses whether to return an adapter, and, if so, chooses according to the provided options.

Called on: GPU this.

Arguments:

Arguments for the GPU.requestAdapter(options) method.

Parameter	Type	Nullable	Optional	Description
options	GPURequestAdapterOptions	✘	✔	Criteria used to select the adapter.

Returns: Promise<GPUAdapter?>

Content timeline steps:

1. Let contentTimeline be the current Content timeline.

2. Let promise be a new promise.

3. Issue the initialization steps on the Device timeline of this.

4. Return promise.

Device timeline initialization steps:

1. All of the requirements in the following steps must be met.

   1. options.featureLevel must be a feature level string.

   If they are met and the user agent chooses to return an adapter:

   1. Set adapter to an adapter chosen according to the rules in § 4.2.2 Adapter Selection and the criteria in options, adhering to § 4.2.1 Adapter Capability Guarantees. Initialize the properties of adapter according to their definitions:

      1. Set adapter.[[limits]] and adapter.[[features]] according to the supported capabilities of the adapter. adapter.[[features]] must contain "core-features-and-limits".

      2. If adapter meets the criteria of a fallback adapter set adapter.[[fallback]] to true. Otherwise, set it to false.

      3. Set adapter.[[xrCompatible]] to options.xrCompatible.

   Otherwise:

   1. Let adapter be null.

2. Issue the subsequent steps on contentTimeline.

Content timeline steps:

1. If adapter is not null:

   1. Resolve promise with a new GPUAdapter encapsulating adapter.

2. Otherwise, Resolve promise with null.

getPreferredCanvasFormat()

Returns an optimal GPUTextureFormat for displaying 8-bit depth, standard dynamic range content on this system. Must only return "rgba8unorm" or "bgra8unorm".

The returned value can be passed as the format to configure() calls on a GPUCanvasContext to ensure the associated canvas is able to display its contents efficiently.

Note: Canvases which are not displayed to the screen may or may not benefit from using this format.

Called on: GPU this.

Returns: GPUTextureFormat

Content timeline steps:

1. Return either "rgba8unorm" or "bgra8unorm", depending on which format is optimal for displaying WebGPU canvases on this system.

GPU has the following attributes:

wgslLanguageFeatures, of type WGSLLanguageFeatures, readonly

The names of supported WGSL language extensions. Supported language extensions are automatically enabled.

Adapters may expire at any time. Upon any change in the system’s state that could affect the result of any requestAdapter() call, the user agent should expire all previously-returned adapters. For example:

• A physical adapter is added/removed (via plug/unplug, driver update, hang recovery, etc.)

• The system’s power configuration has changed (laptop unplugged, power settings changed, etc.)

Note: User agents may choose to expire adapters often, even when there has been no system state change (e.g. seconds or minutes after the adapter was created). This can help obfuscate real system state changes, and make developers more aware that calling requestAdapter() again is always necessary before calling requestDevice(). If an application does encounter this situation, standard device-loss recovery handling should allow it to recover.

const gpuAdapter = await navigator.gpu.requestAdapter();

4.2.1. Adapter Capability Guarantees

Any GPUAdapter returned by requestAdapter() must provide the following guarantees:

• At least one of the following must be true:

  • "texture-compression-bc" is supported.

  • Both "texture-compression-etc2" and "texture-compression-astc" are supported.

• If "texture-compression-bc-sliced-3d" is supported, then "texture-compression-bc" must be supported.

• If "texture-compression-astc-sliced-3d" is supported, then "texture-compression-astc" must be supported.

• All supported limits must be either the default value or better.

• All alignment-class limits must be powers of 2.

• maxBindingsPerBindGroup must be must be ≥ (max bindings per shader stage × max shader stages per pipeline), where:

  • max bindings per shader stage is (maxSampledTexturesPerShaderStage + maxSamplersPerShaderStage + maxStorageBuffersPerShaderStage + maxStorageTexturesPerShaderStage + maxUniformBuffersPerShaderStage).

  • max shader stages per pipeline is 2, because a GPURenderPipeline supports both a vertex and fragment shader.

  Note: maxBindingsPerBindGroup does not reflect a fundamental limit; implementations should raise it to conform to this requirement, rather than lowering the other limits.

• maxBindGroups must be ≤ maxBindGroupsPlusVertexBuffers.

• maxVertexBuffers must be ≤ maxBindGroupsPlusVertexBuffers.

• minUniformBufferOffsetAlignment and minStorageBufferOffsetAlignment must both be ≥ 32 bytes.

  Note: 32 bytes would be the alignment of vec4<f64>. See WebGPU Shading Language § 14.4.1 Alignment and Size.

• maxUniformBufferBindingSize must be ≤ maxBufferSize.

• maxStorageBufferBindingSize must be ≤ maxBufferSize.

• maxStorageBufferBindingSize must be a multiple of 4 bytes.

• maxVertexBufferArrayStride must be a multiple of 4 bytes.

• maxComputeWorkgroupSizeX must be ≤ maxComputeInvocationsPerWorkgroup.

• maxComputeWorkgroupSizeY must be ≤ maxComputeInvocationsPerWorkgroup.

• maxComputeWorkgroupSizeZ must be ≤ maxComputeInvocationsPerWorkgroup.

• maxComputeInvocationsPerWorkgroup must be ≤ maxComputeWorkgroupSizeX × maxComputeWorkgroupSizeY × maxComputeWorkgroupSizeZ.

4.2.2. Adapter Selection

GPURequestAdapterOptions provides hints to the user agent indicating what configuration is suitable for the application.

dictionary GPURequestAdapterOptions {
    DOMString featureLevel = "core";
    GPUPowerPreference powerPreference;
    boolean forceFallbackAdapter = false;
    boolean xrCompatible = false;
};

enum GPUPowerPreference {
    "low-power",
    "high-performance",
};

GPURequestAdapterOptions has the following members:

featureLevel, of type DOMString, defaulting to "core"

"Feature level" for the adapter request.

The allowed feature level string values are:

"core"

No effect.

"compatibility"

No effect.

Note: This value is reserved for future use as a way to opt into additional validation restrictions. Applications should not use this value at this time.

powerPreference, of type GPUPowerPreference

Optionally provides a hint indicating what class of adapter should be selected from the system’s available adapters.

The value of this hint may influence which adapter is chosen, but it must not influence whether an adapter is returned or not.

Note: The primary utility of this hint is to influence which GPU is used in a multi-GPU system. For instance, some laptops have a low-power integrated GPU and a high-performance discrete GPU. This hint may also affect the power configuration of the selected GPU to match the requested power preference.

Note: Depending on the exact hardware configuration, such as battery status and attached displays or removable GPUs, the user agent may select different adapters given the same power preference. Typically, given the same hardware configuration and state and powerPreference, the user agent is likely to select the same adapter.

It must be one of the following values:

undefined (or not present)

Provides no hint to the user agent.

"low-power"

Indicates a request to prioritize power savings over performance.

Note: Generally, content should use this if it is unlikely to be constrained by drawing performance; for example, if it renders only one frame per second, draws only relatively simple geometry with simple shaders, or uses a small HTML canvas element. Developers are encouraged to use this value if their content allows, since it may significantly improve battery life on portable devices.

"high-performance"

Indicates a request to prioritize performance over power consumption.

Note: By choosing this value, developers should be aware that, for devices created on the resulting adapter, user agents are more likely to force device loss, in order to save power by switching to a lower-power adapter. Developers are encouraged to only specify this value if they believe it is absolutely necessary, since it may significantly decrease battery life on portable devices.

forceFallbackAdapter, of type boolean, defaulting to false

When set to true indicates that only a fallback adapter may be returned. If the user agent does not support a fallback adapter, will cause requestAdapter() to resolve to null.

Note: requestAdapter() may still return a fallback adapter if forceFallbackAdapter is set to false and either no other appropriate adapter is available or the user agent chooses to return a fallback adapter. Developers that wish to prevent their applications from running on fallback adapters should check the info.isFallbackAdapter attribute prior to requesting a GPUDevice.

xrCompatible, of type boolean, defaulting to false

When set to true indicates that the best adapter for rendering to a WebXR session must be returned. If the user agent or system does not support WebXR sessions then adapter selection may ignore this value.

Note: If xrCompatible is not set to true when the adapter is requested, GPUDevices created from the adapter cannot be used to render for WebXR sessions.

const gpuAdapter = await navigator.gpu.requestAdapter({
    powerPreference: 'high-performance'
});

4.3. GPUAdapter

A GPUAdapter encapsulates an adapter, and describes its capabilities (features and limits).

To get a GPUAdapter, use requestAdapter().

[Exposed=(Window, Worker), SecureContext]
interface GPUAdapter {
    [SameObject] readonly attribute GPUSupportedFeatures features;
    [SameObject] readonly attribute GPUSupportedLimits limits;
    [SameObject] readonly attribute GPUAdapterInfo info;

    Promise<GPUDevice> requestDevice(optional GPUDeviceDescriptor descriptor = {});
};

GPUAdapter has the following immutable properties

features, of type GPUSupportedFeatures, readonly

The set of values in this.[[adapter]].[[features]].

limits, of type GPUSupportedLimits, readonly

The limits in this.[[adapter]].[[limits]].

info, of type GPUAdapterInfo, readonly

Information about the physical adapter underlying this GPUAdapter.

For a given GPUAdapter, the GPUAdapterInfo values exposed are constant over time.

The same object is returned each time. To create that object for the first time:

Called on: GPUAdapter this.

Returns: GPUAdapterInfo

Content timeline steps:

1. Return a new adapter info for this.[[adapter]].

[[adapter]], of type adapter, readonly

The adapter to which this GPUAdapter refers.

GPUAdapter has the following methods:

requestDevice(descriptor)

Requests a device from the adapter.

This is a one-time action: if a device is returned successfully, the adapter becomes "consumed".

Called on: GPUAdapter this.

Arguments:

Arguments for the GPUAdapter.requestDevice(descriptor) method.

Parameter	Type	Nullable	Optional	Description
descriptor	GPUDeviceDescriptor	✘	✔	Description of the GPUDevice to request.

Returns: Promise<GPUDevice>

Content timeline steps:

1. Let contentTimeline be the current Content timeline.

2. Let promise be a new promise.

3. Let adapter be this.[[adapter]].

4. Issue the initialization steps to the Device timeline of this.

5. Return promise.

Device timeline initialization steps:

1. If any of the following requirements are unmet:

   • The set of values in descriptor.requiredFeatures must be a subset of those in adapter.[[features]].

   Then issue the following steps on contentTimeline and return:

   Content timeline steps:

   1. Reject promise with a TypeError.

   Note: This is the same error that is produced if a feature name isn’t known by the browser at all (in its GPUFeatureName definition). This converges the behavior when the browser doesn’t support a feature with the behavior when a particular adapter doesn’t support a feature.

2. All of the requirements in the following steps must be met.

   1. adapter.[[state]] must not be "consumed".

   2. For each [key, value] in descriptor.requiredLimits for which value is not undefined:

      1. key must be the name of a member of supported limits.

      2. value must be no better than adapter.[[limits]][key].

      3. If key’s class is alignment, value must be a power of 2 less than 2³².

      Note: User agents should consider issuing developer-visible warnings when key is not recognized, even when value is undefined.

   If any are unmet, issue the following steps on contentTimeline and return:

   Content timeline steps:

   1. Reject promise with an OperationError.

3. If adapter.[[state]] is "expired" or the user agent otherwise cannot fulfill the request:

   1. Let device be a new device.

   2. Lose the device(device, "unknown").

   3. Assert adapter.[[state]] is "expired".

      Note: User agents should consider issuing developer-visible warnings in most or all cases when this occurs. Applications should perform reinitialization logic starting with requestAdapter().

   Otherwise:

   1. Let device be a new device with the capabilities described by descriptor.

   2. Expire adapter.

4. Issue the subsequent steps on contentTimeline.

Content timeline steps:

1. Let gpuDevice be a new GPUDevice instance.

2. Set gpuDevice.[[device]] to device.

3. Set device.[[content device]] to gpuDevice.

4. Set gpuDevice.label to descriptor.label.

5. Resolve promise with gpuDevice.

   Note: If the device is already lost because the adapter could not fulfill the request, device.lost has already resolved before promise resolves.

const gpuAdapter = await navigator.gpu.requestAdapter();
const gpuDevice = await gpuAdapter.requestDevice();

4.3.1. GPUDeviceDescriptor

GPUDeviceDescriptor describes a device request.

dictionary GPUDeviceDescriptor
         : GPUObjectDescriptorBase {
    sequence<GPUFeatureName> requiredFeatures = [];
    record<DOMString, (GPUSize64 or undefined)> requiredLimits = {};
    GPUQueueDescriptor defaultQueue = {};
};

GPUDeviceDescriptor has the following members:

requiredFeatures, of type sequence<GPUFeatureName>, defaulting to []

Specifies the features that are required by the device request. The request will fail if the adapter cannot provide these features.

Exactly the specified set of features, and no more or less, will be allowed in validation of API calls on the resulting device.

requiredLimits, of type record<DOMString, (GPUSize64 or undefined)>, defaulting to {}

Specifies the limits that are required by the device request. The request will fail if the adapter cannot provide these limits.

Each key with a non-undefined value must be the name of a member of supported limits.

API calls on the resulting device perform validation according to the exact limits of the device (not the adapter; see § 3.6.2 Limits).

defaultQueue, of type GPUQueueDescriptor, defaulting to {}

The descriptor for the default GPUQueue.

const gpuAdapter = await navigator.gpu.requestAdapter();

const requiredFeatures = [];
if (gpuAdapter.features.has('texture-compression-astc')) {
    requiredFeatures.push('texture-compression-astc')
}

const gpuDevice = await gpuAdapter.requestDevice({
    requiredFeatures
});

const gpuAdapter = await navigator.gpu.requestAdapter();

if (gpuAdapter.limits.maxColorAttachmentBytesPerSample < 64) {
    // When the desired limit isn’t supported, take action to either fall back to a code
    // path that does not require the higher limit or notify the user that their device
    // does not meet minimum requirements.
}

// Request higher limit of max color attachments bytes per sample.
const gpuDevice = await gpuAdapter.requestDevice({
    requiredLimits: { maxColorAttachmentBytesPerSample: 64 },
});

4.3.1.1. GPUFeatureName

Each GPUFeatureName identifies a set of functionality which, if available, allows additional usages of WebGPU that would have otherwise been invalid.

enum GPUFeatureName {
    "core-features-and-limits",
    "depth-clip-control",
    "depth32float-stencil8",
    "texture-compression-bc",
    "texture-compression-bc-sliced-3d",
    "texture-compression-etc2",
    "texture-compression-astc",
    "texture-compression-astc-sliced-3d",
    "timestamp-query",
    "indirect-first-instance",
    "shader-f16",
    "rg11b10ufloat-renderable",
    "bgra8unorm-storage",
    "float32-filterable",
    "float32-blendable",
    "clip-distances",
    "dual-source-blending",
    "subgroups",
    "texture-formats-tier1",
    "texture-formats-tier2",
};

4.4. GPUDevice

A GPUDevice encapsulates a device and exposes the functionality of that device.

GPUDevice is the top-level interface through which WebGPU interfaces are created.

To get a GPUDevice, use requestDevice().

[Exposed=(Window, Worker), SecureContext]
interface GPUDevice : EventTarget {
    [SameObject] readonly attribute GPUSupportedFeatures features;
    [SameObject] readonly attribute GPUSupportedLimits limits;
    [SameObject] readonly attribute GPUAdapterInfo adapterInfo;

    [SameObject] readonly attribute GPUQueue queue;

    undefined destroy();

    GPUBuffer createBuffer(GPUBufferDescriptor descriptor);
    GPUTexture createTexture(GPUTextureDescriptor descriptor);
    GPUSampler createSampler(optional GPUSamplerDescriptor descriptor = {});
    GPUExternalTexture importExternalTexture(GPUExternalTextureDescriptor descriptor);

    GPUBindGroupLayout createBindGroupLayout(GPUBindGroupLayoutDescriptor descriptor);
    GPUPipelineLayout createPipelineLayout(GPUPipelineLayoutDescriptor descriptor);
    GPUBindGroup createBindGroup(GPUBindGroupDescriptor descriptor);

    GPUShaderModule createShaderModule(GPUShaderModuleDescriptor descriptor);
    GPUComputePipeline createComputePipeline(GPUComputePipelineDescriptor descriptor);
    GPURenderPipeline createRenderPipeline(GPURenderPipelineDescriptor descriptor);
    Promise<GPUComputePipeline> createComputePipelineAsync(GPUComputePipelineDescriptor descriptor);
    Promise<GPURenderPipeline> createRenderPipelineAsync(GPURenderPipelineDescriptor descriptor);

    GPUCommandEncoder createCommandEncoder(optional GPUCommandEncoderDescriptor descriptor = {});
    GPURenderBundleEncoder createRenderBundleEncoder(GPURenderBundleEncoderDescriptor descriptor);

    GPUQuerySet createQuerySet(GPUQuerySetDescriptor descriptor);
};
GPUDevice includes GPUObjectBase;

GPUDevice has the following immutable properties:

features, of type GPUSupportedFeatures, readonly

A set containing the GPUFeatureName values of the features supported by the device ([[device]].[[features]]).

limits, of type GPUSupportedLimits, readonly

The limits supported by the device ([[device]].[[limits]]).

queue, of type GPUQueue, readonly

The primary GPUQueue for this device.

adapterInfo, of type GPUAdapterInfo, readonly

Information about the physical adapter which created the device that this GPUDevice refers to.

For a given GPUDevice, the GPUAdapterInfo values exposed are constant over time.

The same object is returned each time. To create that object for the first time:

Called on: GPUDevice this.

Returns: GPUAdapterInfo

Content timeline steps:

1. Return a new adapter info for this.[[device]].[[adapter]].

The [[device]] for a GPUDevice is the device that the GPUDevice refers to.

GPUDevice has the following methods:

destroy()

Destroys the device, preventing further operations on it. Outstanding asynchronous operations will fail.

Note: It is valid to destroy a device multiple times.

Called on: GPUDevice this.

Content timeline steps:

1. unmap() all GPUBuffers from this device.

2. Issue the subsequent steps on the Device timeline of this.

1. Lose the device(this.[[device]], "destroyed").

Note: Since no further operations can be enqueued on this device, implementations can abort outstanding asynchronous operations immediately and free resource allocations, including mapped memory that was just unmapped.

4.5. Example

let gpuDevice = null;

async function initializeWebGPU() {
    // Check to ensure the user agent supports WebGPU.
    if (!('gpu' in navigator)) {
        console.error("User agent doesn’t support WebGPU.");
        return false;
    }

    // Request an adapter.
    const gpuAdapter = await navigator.gpu.requestAdapter();

    // requestAdapter may resolve with null if no suitable adapters are found.
    if (!gpuAdapter) {
        console.error('No WebGPU adapters found.');
        return false;
    }

    // Request a device.
    // Note that the promise will reject if invalid options are passed to the optional
    // dictionary. To avoid the promise rejecting always check any features and limits
    // against the adapters features and limits prior to calling requestDevice().
    gpuDevice = await gpuAdapter.requestDevice();

    // requestDevice will never return null, but if a valid device request can’t be
    // fulfilled for some reason it may resolve to a device which has already been lost.
    // Additionally, devices can be lost at any time after creation for a variety of reasons
    // (ie: browser resource management, driver updates), so it’s a good idea to always
    // handle lost devices gracefully.
    gpuDevice.lost.then((info) => {
        console.error(`WebGPU device was lost: ${info.message}`);

        gpuDevice = null;

        // Many causes for lost devices are transient, so applications should try getting a
        // new device once a previous one has been lost unless the loss was caused by the
        // application intentionally destroying the device. Note that any WebGPU resources
        // created with the previous device (buffers, textures, etc) will need to be
        // re-created with the new one.
        if (info.reason != 'destroyed') {
            initializeWebGPU();
        }
    });

    onWebGPUInitialized();

    return true;
}

function onWebGPUInitialized() {
    // Begin creating WebGPU resources here...
}

initializeWebGPU();

5. Buffers

5.1. GPUBuffer

A GPUBuffer represents a block of memory that can be used in GPU operations. Data is stored in linear layout, meaning that each byte of the allocation can be addressed by its offset from the start of the GPUBuffer, subject to alignment restrictions depending on the operation. Some GPUBuffers can be mapped which makes the block of memory accessible via an ArrayBuffer called its mapping.

GPUBuffers are created via createBuffer(). Buffers may be mappedAtCreation.

[Exposed=(Window, Worker), SecureContext]
interface GPUBuffer {
    readonly attribute GPUSize64Out size;
    readonly attribute GPUFlagsConstant usage;

    readonly attribute GPUBufferMapState mapState;

    Promise<undefined> mapAsync(GPUMapModeFlags mode, optional GPUSize64 offset = 0, optional GPUSize64 size);
    ArrayBuffer getMappedRange(optional GPUSize64 offset = 0, optional GPUSize64 size);
    undefined unmap();

    undefined destroy();
};
GPUBuffer includes GPUObjectBase;

enum GPUBufferMapState {
    "unmapped",
    "pending",
    "mapped",
};

GPUBuffer has the following immutable properties:

size, of type GPUSize64Out, readonly

The length of the GPUBuffer allocation in bytes.

usage, of type GPUFlagsConstant, readonly

The allowed usages for this GPUBuffer.

GPUBuffer has the following content timeline properties:

mapState, of type GPUBufferMapState, readonly

The current GPUBufferMapState of the buffer:

"unmapped"

The buffer is not mapped for use by this.getMappedRange().

"pending"

A mapping of the buffer has been requested, but is pending. It may succeed, or fail validation in mapAsync().

"mapped"

The buffer is mapped and this.getMappedRange() may be used.

The getter steps are:

Content timeline steps:

1. If this.[[mapping]] is not null, return "mapped".

2. If this.[[pending_map]] is not null, return "pending".

3. Return "unmapped".

[[pending_map]], of type Promise<void> or null, initially null

The Promise returned by the currently-pending mapAsync() call.

There is never more than one pending map, because mapAsync() will refuse immediately if a request is already in flight.

[[mapping]], of type active buffer mapping or null, initially null

Set if and only if the buffer is currently mapped for use by getMappedRange(). Null otherwise (even if there is a [[pending_map]]).

An active buffer mapping is a structure with the following fields:

data, of type Data Block

The mapping for this GPUBuffer. This data is accessed through ArrayBuffers which are views onto this data, returned by getMappedRange() and stored in views.

mode, of type GPUMapModeFlags

The GPUMapModeFlags of the map, as specified in the corresponding call to mapAsync() or createBuffer().

range, of type tuple [unsigned long long, unsigned long long]

The range of this GPUBuffer that is mapped.

views, of type list<ArrayBuffer>

The ArrayBuffers returned via getMappedRange() to the application. They are tracked so they can be detached when unmap() is called.

To initialize an active buffer mapping with mode mode and range range, run the following content timeline steps:

1. Let size be range[1] - range[0].

2. Let data be ? CreateByteDataBlock(size).

   NOTE:

   This may result in a RangeError being thrown. For consistency and predictability:

   • For any size at which new ArrayBuffer() would succeed at a given moment, this allocation should succeed at that moment.

   • For any size at which new ArrayBuffer() deterministically throws a RangeError, this allocation should as well.

3. Return an active buffer mapping with:

   • data set to data.

   • mode set to mode.

   • range set to range.

   • views set to [].

GPUBuffer has the following device timeline properties:

[[internal state]]

The current internal state of the buffer:

"available"

The buffer may be used in queue operations (unless it is invalid).

"unavailable"

The buffer may not be used in queue operations due to being mapped.

"destroyed"

The buffer may not be used in any operations due to being destroy()ed.

5.1.1. GPUBufferDescriptor

dictionary GPUBufferDescriptor
         : GPUObjectDescriptorBase {
    required GPUSize64 size;
    required GPUBufferUsageFlags usage;
    boolean mappedAtCreation = false;
};

GPUBufferDescriptor has the following members:

size, of type GPUSize64

The size of the buffer in bytes.

usage, of type GPUBufferUsageFlags

The allowed usages for the buffer.

mappedAtCreation, of type boolean, defaulting to false

If true creates the buffer in an already mapped state, allowing getMappedRange() to be called immediately. It is valid to set mappedAtCreation to true even if usage does not contain MAP_READ or MAP_WRITE. This can be used to set the buffer’s initial data.

Guarantees that even if the buffer creation eventually fails, it will still appear as if the mapped range can be written/read to until it is unmapped.

5.1.2. Buffer Usages

typedef [EnforceRange] unsigned long GPUBufferUsageFlags;
[Exposed=(Window, Worker), SecureContext]
namespace GPUBufferUsage {
    const GPUFlagsConstant MAP_READ      = 0x0001;
    const GPUFlagsConstant MAP_WRITE     = 0x0002;
    const GPUFlagsConstant COPY_SRC      = 0x0004;
    const GPUFlagsConstant COPY_DST      = 0x0008;
    const GPUFlagsConstant INDEX         = 0x0010;
    const GPUFlagsConstant VERTEX        = 0x0020;
    const GPUFlagsConstant UNIFORM       = 0x0040;
    const GPUFlagsConstant STORAGE       = 0x0080;
    const GPUFlagsConstant INDIRECT      = 0x0100;
    const GPUFlagsConstant QUERY_RESOLVE = 0x0200;
};

The GPUBufferUsage flags determine how a GPUBuffer may be used after its creation:

MAP_READ

The buffer can be mapped for reading. (Example: calling mapAsync() with GPUMapMode.READ)

May only be combined with COPY_DST.

MAP_WRITE

The buffer can be mapped for writing. (Example: calling mapAsync() with GPUMapMode.WRITE)

May only be combined with COPY_SRC.

COPY_SRC

The buffer can be used as the source of a copy operation. (Examples: as the source argument of a copyBufferToBuffer() or copyBufferToTexture() call.)

COPY_DST

The buffer can be used as the destination of a copy or write operation. (Examples: as the destination argument of a copyBufferToBuffer() or copyTextureToBuffer() call, or as the target of a writeBuffer() call.)

INDEX

The buffer can be used as an index buffer. (Example: passed to setIndexBuffer().)

VERTEX

The buffer can be used as a vertex buffer. (Example: passed to setVertexBuffer().)

UNIFORM

The buffer can be used as a uniform buffer. (Example: as a bind group entry for a GPUBufferBindingLayout with a buffer.type of "uniform".)

STORAGE

The buffer can be used as a storage buffer. (Example: as a bind group entry for a GPUBufferBindingLayout with a buffer.type of "storage" or "read-only-storage".)

INDIRECT

The buffer can be used as to store indirect command arguments. (Examples: as the indirectBuffer argument of a drawIndirect() or dispatchWorkgroupsIndirect() call.)

QUERY_RESOLVE

The buffer can be used to capture query results. (Example: as the destination argument of a resolveQuerySet() call.)

5.1.3. Buffer Creation

createBuffer(descriptor)

Creates a GPUBuffer.

Called on: GPUDevice this.

Arguments:

Arguments for the GPUDevice.createBuffer(descriptor) method.

Parameter	Type	Nullable	Optional	Description
descriptor	GPUBufferDescriptor	✘	✘	Description of the GPUBuffer to create.

Returns: GPUBuffer

Content timeline steps:

1. Let b be ! create a new WebGPU object(this, GPUBuffer, descriptor).

2. Set b.size to descriptor.size.

3. Set b.usage to descriptor.usage.

4. If descriptor.mappedAtCreation is true:

   1. If descriptor.size is not a multiple of 4, throw a RangeError.

   2. Set b.[[mapping]] to ? initialize an active buffer mapping with mode WRITE and range [0, descriptor.size].

5. Issue the initialization steps on the Device timeline of this.

6. Return b.

Device timeline initialization steps:

1. If any of the following requirements are unmet, generate a validation error, invalidate b and return.

   • this must not be lost.

   • descriptor.usage must not be 0.

   • descriptor.usage must be a subset of the allowed buffer usages for this.

   • If descriptor.usage contains MAP_READ:

     • descriptor.usage must contain no other flags except COPY_DST.

   • If descriptor.usage contains MAP_WRITE:

     • descriptor.usage must contain no other flags except COPY_SRC.

   • If descriptor.size must be ≤ this.[[device]].[[limits]].maxBufferSize.

Note: If buffer creation fails, and descriptor.mappedAtCreation is false, any calls to mapAsync() will reject, so any resources allocated to enable mapping can and may be discarded or recycled.

1. If descriptor.mappedAtCreation is true:

   1. Set b.[[internal state]] to "unavailable".

   Else:

   1. Set b.[[internal state]] to "available".

2. Create a device allocation for b where each byte is zero.

   If the allocation fails without side-effects, generate an out-of-memory error, invalidate b, and return.

const buffer = gpuDevice.createBuffer({
    size: 128,
    usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST
});

5.1.4. Buffer Destruction

An application that no longer requires a GPUBuffer can choose to lose access to it before garbage collection by calling destroy(). Destroying a buffer also unmaps it, freeing any memory allocated for the mapping.

Note: This allows the user agent to reclaim the GPU memory associated with the GPUBuffer once all previously submitted operations using it are complete.

GPUBuffer has the following methods:

destroy()

Destroys the GPUBuffer.

Note: It is valid to destroy a buffer multiple times.

Called on: GPUBuffer this.

Returns: undefined

Content timeline steps:

1. Call this.unmap().

2. Issue the subsequent steps on the Device timeline of this.[[device]].

Device timeline steps:

1. Set this.[[internal state]] to "destroyed".

Note: Since no further operations can be enqueued using this buffer, implementations can free resource allocations, including mapped memory that was just unmapped.

5.2. Buffer Mapping

An application can request to map a GPUBuffer so that they can access its content via ArrayBuffers that represent part of the GPUBuffer’s allocations. Mapping a GPUBuffer is requested asynchronously with mapAsync() so that the user agent can ensure the GPU finished using the GPUBuffer before the application can access its content. A mapped GPUBuffer cannot be used by the GPU and must be unmapped using unmap() before work using it can be submitted to the Queue timeline.

Once the GPUBuffer is mapped, the application can synchronously ask for access to ranges of its content with getMappedRange(). The returned ArrayBuffer can only be detached by unmap() (directly, or via GPUBuffer.destroy() or GPUDevice.destroy()), and cannot be transferred. A TypeError is thrown by any other operation that attempts to do so.

typedef [EnforceRange] unsigned long GPUMapModeFlags;
[Exposed=(Window, Worker), SecureContext]
namespace GPUMapMode {
    const GPUFlagsConstant READ  = 0x0001;
    const GPUFlagsConstant WRITE = 0x0002;
};

The GPUMapMode flags determine how a GPUBuffer is mapped when calling mapAsync():

READ

Only valid with buffers created with the MAP_READ usage.

Once the buffer is mapped, calls to getMappedRange() will return an ArrayBuffer containing the buffer’s current values. Changes to the returned ArrayBuffer will be discarded after unmap() is called.

WRITE

Only valid with buffers created with the MAP_WRITE usage.

Once the buffer is mapped, calls to getMappedRange() will return an ArrayBuffer containing the buffer’s current values. Changes to the returned ArrayBuffer will be stored in the GPUBuffer after unmap() is called.

Note: Since the MAP_WRITE buffer usage may only be combined with the COPY_SRC buffer usage, mapping for writing can never return values produced by the GPU, and the returned ArrayBuffer will only ever contain the default initialized data (zeros) or data written by the webpage during a previous mapping.

GPUBuffer has the following methods:

mapAsync(mode, offset, size)

Maps the given range of the GPUBuffer and resolves the returned Promise when the GPUBuffer’s content is ready to be accessed with getMappedRange().

The resolution of the returned Promise only indicates that the buffer has been mapped. It does not guarantee the completion of any other operations visible to the content timeline, and in particular does not imply that any other Promise returned from onSubmittedWorkDone() or mapAsync() on other GPUBuffers have resolved.

The resolution of the Promise returned from onSubmittedWorkDone() does imply the completion of mapAsync() calls made prior to that call, on GPUBuffers last used exclusively on that queue.

Called on: GPUBuffer this.

Arguments:

Arguments for the GPUBuffer.mapAsync(mode, offset, size) method.

Parameter	Type	Nullable	Optional	Description
mode	GPUMapModeFlags	✘	✘	Whether the buffer should be mapped for reading or writing.
offset	GPUSize64	✘	✔	Offset in bytes into the buffer to the start of the range to map.
size	GPUSize64	✘	✔	Size in bytes of the range to map.

Returns: Promise<undefined>

Content timeline steps:

1. Let contentTimeline be the current Content timeline.

2. If this.mapState is not "unmapped":

   1. Issue the early-reject steps on the Device timeline of this.[[device]].

   2. Return a promise rejected with OperationError.

3. Let p be a new Promise.

4. Set this.[[pending_map]] to p.

5. Issue the validation steps on the Device timeline of this.[[device]].

6. Return p.

Device timeline early-reject steps:

1. Generate a validation error.

2. Return.

Device timeline validation steps:

1. If size is undefined:

   1. Let rangeSize be max(0, this.size - offset).

   Otherwise:

   1. Let rangeSize be size.

2. If any of the following conditions are unsatisfied:

   • this must be valid.

   1. Set deviceLost to true.

   2. Issue the map failure steps on contentTimeline.

   3. Return.

3. If any of the following conditions are unsatisfied:

   • this.[[internal state]] is "available".

   • offset is a multiple of 8.

   • rangeSize is a multiple of 4.

   • offset + rangeSize ≤ this.size

   • mode contains only bits defined in GPUMapMode.

   • mode contains exactly one of READ or WRITE.

   • If mode contains READ then this.usage must contain MAP_READ.

   • If mode contains WRITE then this.usage must contain MAP_WRITE.

   Then:

   1. Set deviceLost to false.

   2. Issue the map failure steps on contentTimeline.

   3. Generate a validation error.

   4. Return.

4. Set this.[[internal state]] to "unavailable".

   Note: Since the buffer is mapped, its contents cannot change between this step and unmap().

5. When either of the following events occur (whichever comes first), or if either has already occurred:

   • The device timeline becomes informed of the completion of an unspecified queue timeline point:

     • after the completion of currently-enqueued operations that use this

     • and no later than the completion of all currently-enqueued operations (regardless of whether they use this).

   • this.[[device]] becomes lost.

   Then issue the subsequent steps on the device timeline of this.[[device]].

Device timeline steps:

1. Set deviceLost to true if this.[[device]] is lost, and false otherwise.

   Note: The device could have been lost between the previous block of steps and this one.

2. If deviceLost:

   1. Issue the map failure steps on contentTimeline.

   Otherwise:

   1. Let internalStateAtCompletion be this.[[internal state]].

      Note: If, and only if, at this point the buffer has become "available" again due to an unmap() call, then [[pending_map]] != p below, so mapping will not succeed in the steps below.

   2. Let dataForMappedRegion be the contents of this starting at offset offset, for rangeSize bytes.

   3. Issue the map success steps on the contentTimeline.

Content timeline map success steps:

1. If this.[[pending_map]] != p:

   Note: The map has been cancelled by unmap().

   1. Assert p is rejected.

   2. Return.

2. Assert p is pending.

3. Assert internalStateAtCompletion is "unavailable".

4. Let mapping be initialize an active buffer mapping with mode mode and range [offset, offset + rangeSize].

   If this allocation fails:

   1. Set this.[[pending_map]] to null, and reject p with a RangeError.

   2. Return.

5. Set the content of mapping.data to dataForMappedRegion.

6. Set this.[[mapping]] to mapping.

7. Set this.[[pending_map]] to null, and resolve p.

Content timeline map failure steps:

1. If this.[[pending_map]] != p:

   Note: The map has been cancelled by unmap().

   1. Assert p is already rejected.

   2. Return.

2. Assert p is still pending.

3. Set this.[[pending_map]] to null.

4. If deviceLost:

   1. Reject p with an AbortError.

      Note: This is the same error type produced by cancelling the map using unmap().

   Otherwise:

   1. Reject p with an OperationError.

getMappedRange(offset, size)

Returns an ArrayBuffer with the contents of the GPUBuffer in the given mapped range.

Called on: GPUBuffer this.

Arguments:

Arguments for the GPUBuffer.getMappedRange(offset, size) method.

Parameter	Type	Nullable	Optional	Description
offset	GPUSize64	✘	✔	Offset in bytes into the buffer to return buffer contents from.
size	GPUSize64	✘	✔	Size in bytes of the ArrayBuffer to return.

Returns: ArrayBuffer

Content timeline steps:

1. If size is missing:

   1. Let rangeSize be max(0, this.size - offset).

   Otherwise, let rangeSize be size.

2. If any of the following conditions are unsatisfied, throw an OperationError and return.

   • this.[[mapping]] is not null.

   • offset is a multiple of 8.

   • rangeSize is a multiple of 4.

   • offset ≥ this.[[mapping]].range[0].

   • offset + rangeSize ≤ this.[[mapping]].range[1].

   • [offset, offset + rangeSize) does not overlap another range in this.[[mapping]].views.

   Note: It is always valid to get mapped ranges of a GPUBuffer that is mappedAtCreation, even if it is invalid, because the Content timeline might not know it is invalid.

3. Let data be this.[[mapping]].data.

4. Let view be ! create an ArrayBuffer of size rangeSize, but with its pointer mutably referencing the content of data at offset (offset - [[mapping]].range[0]).

   Note: A RangeError may not be thrown here, because the data has already been allocated during mapAsync() or createBuffer().

5. Set view.[[ArrayBufferDetachKey]] to "WebGPUBufferMapping".

   Note: This causes a TypeError to be thrown if an attempt is made to DetachArrayBuffer, except by unmap().

6. Append view to this.[[mapping]].views.

7. Return view.

Note: User agents should consider issuing a developer-visible warning if getMappedRange() succeeds without having checked the status of the map, by waiting for mapAsync() to succeed, querying a mapState of "mapped", or waiting for a later onSubmittedWorkDone() call to succeed.

unmap()

Unmaps the mapped range of the GPUBuffer and makes its contents available for use by the GPU again.

Called on: GPUBuffer this.

Returns: undefined

Content timeline steps:

1. If this.[[pending_map]] is not null:

   1. Reject this.[[pending_map]] with an AbortError.

   2. Set this.[[pending_map]] to null.

2. If this.[[mapping]] is null:

   1. Return.

3. For each ArrayBuffer ab in this.[[mapping]].views:

   1. Perform DetachArrayBuffer(ab, "WebGPUBufferMapping").

4. Let bufferUpdate be null.

5. If this.[[mapping]].mode contains WRITE:

   1. Set bufferUpdate to { data: this.[[mapping]].data, offset: this.[[mapping]].range[0] }.

   Note: When a buffer is mapped without the WRITE mode, then unmapped, any local modifications done by the application to the mapped ranges ArrayBuffer are discarded and will not affect the content of later mappings.

6. Set this.[[mapping]] to null.

7. Issue the subsequent steps on the Device timeline of this.[[device]].

Device timeline steps:

1. If any of the following conditions are unsatisfied, return.

   • this is valid to use with this.[[device]].

2. Assert this.[[internal state]] is "unavailable".

3. If bufferUpdate is not null:

   1. Issue the following steps on the Queue timeline of this.[[device]].queue:

      Queue timeline steps:

      1. Update the contents of this at offset bufferUpdate.offset with the data bufferUpdate.data.

4. Set this.[[internal state]] to "available".

6. Textures and Texture Views

6.1. GPUTexture

A texture is made up of 1d, 2d, or 3d arrays of data which can contain multiple values per-element to represent things like colors. Textures can be read and written in many ways, depending on the GPUTextureUsage they are created with. For example, textures can be sampled, read, and written from render and compute pipeline shaders, and they can be written by render pass outputs. Internally, textures are often stored in GPU memory with a layout optimized for multidimensional access rather than linear access.

One texture consists of one or more texture subresources, each uniquely identified by a mipmap level and, for 2d textures only, array layer and aspect.

A texture subresource is a subresource: each can be used in different internal usages within a single usage scope.

Each subresource in a mipmap level is approximately half the size, in each spatial dimension, of the corresponding resource in the lesser level (see logical miplevel-specific texture extent). The subresource in level 0 has the dimensions of the texture itself. Smaller levels are typically used to store lower resolution versions of the same image. GPUSampler and WGSL provide facilities for selecting and interpolating between levels of detail, explicitly or automatically.

A "2d" texture may be an array of array layers. Each subresource in a layer is the same size as the corresponding resources in other layers. For non-2d textures, all subresources have an array layer index of 0.

Each subresource has an aspect. Color textures have just one aspect: color. Depth-or-stencil format textures may have multiple aspects: a depth aspect, a stencil aspect, or both, and may be used in special ways, such as in depthStencilAttachment and in "depth" bindings.

A "3d" texture may have multiple slices, each being the two-dimensional image at a particular z value in the texture. Slices are not separate subresources.

[Exposed=(Window, Worker), SecureContext]
interface GPUTexture {
    GPUTextureView createView(optional GPUTextureViewDescriptor descriptor = {});

    undefined destroy();

    readonly attribute GPUIntegerCoordinateOut width;
    readonly attribute GPUIntegerCoordinateOut height;
    readonly attribute GPUIntegerCoordinateOut depthOrArrayLayers;
    readonly attribute GPUIntegerCoordinateOut mipLevelCount;
    readonly attribute GPUSize32Out sampleCount;
    readonly attribute GPUTextureDimension dimension;
    readonly attribute GPUTextureFormat format;
    readonly attribute GPUFlagsConstant usage;
};
GPUTexture includes GPUObjectBase;

GPUTexture has the following immutable properties:

width, of type GPUIntegerCoordinateOut, readonly

The width of this GPUTexture.

height, of type GPUIntegerCoordinateOut, readonly

The height of this GPUTexture.

depthOrArrayLayers, of type GPUIntegerCoordinateOut, readonly

The depth or layer count of this GPUTexture.

mipLevelCount, of type GPUIntegerCoordinateOut, readonly

The number of mip levels of this GPUTexture.

sampleCount, of type GPUSize32Out, readonly

The number of sample count of this GPUTexture.

dimension, of type GPUTextureDimension, readonly

The dimension of the set of texel for each of this GPUTexture’s subresources.

format, of type GPUTextureFormat, readonly

The format of this GPUTexture.

usage, of type GPUFlagsConstant, readonly

The allowed usages for this GPUTexture.

[[viewFormats]], of type sequence<GPUTextureFormat>

The set of GPUTextureFormats that can be used as the GPUTextureViewDescriptor.format when creating views on this GPUTexture.

GPUTexture has the following device timeline properties:

[[destroyed]], of type boolean, initially false

If the texture is destroyed, it can no longer be used in any operation, and its underlying memory can be freed.

The logical miplevel-specific texture extent of a texture is the size of the texture in texels at a specific miplevel. It is calculated by this procedure:

The physical miplevel-specific texture extent of a texture is the size of the texture in texels at a specific miplevel that includes the possible extra padding to form complete texel blocks in the texture. It is calculated by this procedure:

6.1.1. GPUTextureDescriptor

dictionary GPUTextureDescriptor
         : GPUObjectDescriptorBase {
    required GPUExtent3D size;
    GPUIntegerCoordinate mipLevelCount = 1;
    GPUSize32 sampleCount = 1;
    GPUTextureDimension dimension = "2d";
    required GPUTextureFormat format;
    required GPUTextureUsageFlags usage;
    sequence<GPUTextureFormat> viewFormats = [];
};

GPUTextureDescriptor has the following members:

size, of type GPUExtent3D

The width, height, and depth or layer count of the texture.

mipLevelCount, of type GPUIntegerCoordinate, defaulting to 1

The number of mip levels the texture will contain.

sampleCount, of type GPUSize32, defaulting to 1

The sample count of the texture. A sampleCount > 1 indicates a multisampled texture.

dimension, of type GPUTextureDimension, defaulting to "2d"

Whether the texture is one-dimensional, an array of two-dimensional layers, or three-dimensional.

format, of type GPUTextureFormat

The format of the texture.

usage, of type GPUTextureUsageFlags

The allowed usages for the texture.

viewFormats, of type sequence<GPUTextureFormat>, defaulting to []

Specifies what view format values will be allowed when calling createView() on this texture (in addition to the texture’s actual format).

NOTE:

Adding a format to this list may have a significant performance impact, so it is best to avoid adding formats unnecessarily.

The actual performance impact is highly dependent on the target system; developers must test various systems to find out the impact on their particular application. For example, on some systems any texture with a format or viewFormats entry including "rgba8unorm-srgb" will perform less optimally than a "rgba8unorm" texture which does not. Similar caveats exist for other formats and pairs of formats on other systems.

Formats in this list must be texture view format compatible with the texture format.

Two GPUTextureFormats format and viewFormat are texture view format compatible if:

• format equals viewFormat, or

• format and viewFormat differ only in whether they are srgb formats (have the -srgb suffix).

enum GPUTextureDimension {
    "1d",
    "2d",
    "3d",
};

"1d"

Specifies a texture that has one dimension, width. "1d" textures cannot have mipmaps, be multisampled, use compressed or depth/stencil formats, or be used as a render target.

"2d"

Specifies a texture that has a width and height, and may have layers.

"3d"

Specifies a texture that has a width, height, and depth. "3d" textures cannot be multisampled, and their format must support 3d textures (all plain color formats and some packed/compressed formats).

6.1.2. Texture Usages

typedef [EnforceRange] unsigned long GPUTextureUsageFlags;
[Exposed=(Window, Worker), SecureContext]
namespace GPUTextureUsage {
    const GPUFlagsConstant COPY_SRC          = 0x01;
    const GPUFlagsConstant COPY_DST          = 0x02;
    const GPUFlagsConstant TEXTURE_BINDING   = 0x04;
    const GPUFlagsConstant STORAGE_BINDING   = 0x08;
    const GPUFlagsConstant RENDER_ATTACHMENT = 0x10;
};

The GPUTextureUsage flags determine how a GPUTexture may be used after its creation:

COPY_SRC

The texture can be used as the source of a copy operation. (Examples: as the source argument of a copyTextureToTexture() or copyTextureToBuffer() call.)

COPY_DST

The texture can be used as the destination of a copy or write operation. (Examples: as the destination argument of a copyTextureToTexture() or copyBufferToTexture() call, or as the target of a writeTexture() call.)

TEXTURE_BINDING

The texture can be bound for use as a sampled texture in a shader (Example: as a bind group entry for a GPUTextureBindingLayout.)

STORAGE_BINDING

The texture can be bound for use as a storage texture in a shader (Example: as a bind group entry for a GPUStorageTextureBindingLayout.)

RENDER_ATTACHMENT

The texture can be used as a color or depth/stencil attachment in a render pass. (Example: as a GPURenderPassColorAttachment.view or GPURenderPassDepthStencilAttachment.view.)

6.1.3. Texture Creation

createTexture(descriptor)

Creates a GPUTexture.

Called on: GPUDevice this.

Arguments:

Arguments for the GPUDevice.createTexture(descriptor) method.

Parameter	Type	Nullable	Optional	Description
descriptor	GPUTextureDescriptor	✘	✘	Description of the GPUTexture to create.

Returns: GPUTexture

Content timeline steps:

1. ? validate GPUExtent3D shape(descriptor.size).

2. ? Validate texture format required features of descriptor.format with this.[[device]].

3. ? Validate texture format required features of each element of descriptor.viewFormats with this.[[device]].

4. Let t be ! create a new WebGPU object(this, GPUTexture, descriptor).

5. Set t.width to descriptor.size.width.

6. Set t.height to descriptor.size.height.

7. Set t.depthOrArrayLayers to descriptor.size.depthOrArrayLayers.

8. Set t.mipLevelCount to descriptor.mipLevelCount.

9. Set t.sampleCount to descriptor.sampleCount.

10. Set t.dimension to descriptor.dimension.

11. Set t.format to descriptor.format.

12. Set t.usage to descriptor.usage.

13. Issue the initialization steps on the Device timeline of this.

14. Return t.

Device timeline initialization steps:

1. If any of the following conditions are unsatisfied generate a validation error, invalidate t and return.

   • validating GPUTextureDescriptor(this, descriptor) returns true.

2. Set t.[[viewFormats]] to descriptor.viewFormats.

3. Create a device allocation for t where each block has an equivalent texel representation to a block with a bit representation of zero.

   If the allocation fails without side-effects, generate an out-of-memory error, invalidate t, and return.